Independent machine-learning R&D by Luca Finn Law.
Germany · open source · two projects
Prototype · live demo
A 2.2M-parameter model that beats a perfect detector
Vision pre-shapes the grip before contact.
Active research
Five local models, and a result we withdrew
Japanese-to-English literary translation, run locally.
Five stages on a single consumer GPU, fully offline, EPUB in and out. The current fine-tune measures below its own base model. An earlier positive result was traced to an answer-key leak and retracted.
Models trained on hardware here.
Inference on the device.
Results published including the failures.